Efficient inference, agent behavior, and rigorous evaluation AI is about the workflow, not the model. AMD Halo X Group Chat: https://xitter.cf/t.co/NqMCJamWhI
United States
Joined May 2009
- Tweets30.5K
- Following979
- Followers7.9K
- Likes81.2K
Pinned Tweet
64 agents. Two AMD Strix Halo desktops. Everything running local.
32 builds per machine, all at the same time. Games, apps, tools, websites. Each one generating its own working HTML/CSS/JS project.
64 apps finished in 5:32.
1,166 tok/s combined peak
662 tok/s combined average
~220,000 output tokens
All from two Ryzen AI MAX+ 395 systems running my optimized Gemma 4 26B stack.
And these aren't screenshots. You can open the wall and actually use what they built:
llm.ciru.ai/research/amd-64-…
AMD provided one of the two Ryzen AI Halo systems used for this demo.
Try the custom Ciru model for yourself:
huggingface.co/jcbtc/Gemma4-…
Ciru.ai - Crown 👑 retweeted
If you think that DSV4.1 to V4.1-1017 improvement is solely due to OpenAI distillation, then I wonder why OpenAI couldn't distill a Luna that beats even V4.1.
Just how tiny must Luna be then? and how high their margins, given that it's slower too, despite superior hardware?
Ciru.ai - Crown 👑 retweeted
Wheres @theo to tell me why this is garbage?
🚀 DeepSeek-V4.1-Flash-1017 is here!
Stronger benchmark performance, faster task completion, and shorter outputs than the previous version—a Pareto improvement in performance and cost. Now live on our website, app, and API.
🛠️ Optimized for @DeepSeekHarness, Tencent CodeBuddy, WorkBuddy, and more. Upgrade to DeepSeek Harness v0.2.1-rc.1 or later for the best experience with the new model.
1/4
Ciru.ai - Crown 👑 retweeted
Imagine flying a roadster at 200 mph and talking to Grok bot to kick off a task with Opus to build something and then merge it when CI turns green. The future is here guys.
I love this.
Memory management is one of the largest problems with agentic fleets. Tencent released a model trained specifically for memory management, with their own stack.
I definitely want to try this out. I've been looking for memory solutions for a while, it's not easy.
@bradmillscan
A 3B active model beats Opus 5 at memory. We open-sourced the full stack — model, engine, and app.
Hy Memory handles chat, work, and code memory in one layer, and it's first on all four benchmarks we ran, leading MemoryArena by 10.2 points while cutting memory volume 51.8% and recall tokens 39.3%.
The model isn't tied to our engine. Drop Hy-Memory-30B-A3B into Mem0, MemOS, or EverMemOS and it still comes out on top, even on the one system it never trained on. It's already running in our own products, where one assistant's multi-hop recall went from 50% to 90%. It also unifies memory across Codex, Cursor, WorkBuddy, and more.
Repo: github.com/Tencent-Hunyuan/H…
Weights: huggingface.co/tencent/Hy-Me…
Questions welcome in the replies.
I can't stand this and it makes me so happy to be using models locally.
Claude Opus 5.5 got nerfed!
I revisited my original benchmark today, using the same repository, the same 9 planted defects, and the same open-ended coding challenge.
Previous results:
Sep 22: 9/9 defects addressed
Oct 3/4: 9/9 defects addressed
Today (Oct 15):
Opus 5.5 in Cursor: 7/9 in 9 minutes
Opus 5.5 in Claude Code: 7/9 in 12m 50s
Both preserved the GT-10 restraint condition, but both left two defect signatures untouched.
Interestingly, they missed different defects despite using the same underlying model.
Both runs passed their test suites and confidently declared the repository ready to ship.
That's what caught my attention: the agents decided they were done while the hidden checks still flagged unresolved issues.
And this wasn't a one-off. I repeated the experiment 3 times in each harness. Same result.
That's 6 runs across two different coding environments.
The results strongly suggest a performance regression on this benchmark, although the signature checks alone can't prove the model itself was downgraded.
Going from 9/9 to 7/9 consistently is disappointing.
Has anyone else noticed Opus 5.5 performing worse lately?
A treat for those of you who like high precision over pure speed. I've created a new custom build with @OpenBMB's excellent MiniCPM4.7 35b model!
Hadamard-rotated 8-bit expert weights, with 256-weight groups and FP16 scales. The remaining weights keeping their original precision, mostly BF16
I also included several optimizations to make it very good at concurrent loads. The headline TPS numbers might not be as impressive as a Q4, but it's so solid that it ends up getting more complex workflows finished faster. This model completes my previous 8 agent 3D AMD Logo demo in just 2:20, compared to nearly 11 minutes with the previous Q4 35b model I used.
40.45 generated tokens/s with a 63,870-token prompt with 1,968.6 prompt tokens/s.
122.78 aggregate generated tokens/s across eight concurrent vision/tool requests, with thinking enabled.
262,144-token maximum context, with the headline long-context speed measurement taken at approximately 64K, and an 804k token context pool to share.
Some of the kernel improvements were substantial. The retained attention change reduced measured component latency by approximately 41–50%, with bitwise-identical outputs on captured operands. The convolution traversal change reduced measured component latency by approximately 86–87%, with live output and state comparisons remaining bitwise identical.
I also checked numerical fidelity against the original BF16 model. Matched-prefix testing measured 99.23% top-token agreement on live single-request continuations and 99.21% on live eight-request continuations. A separate broader frozen-prefill panel measured 96.67%.
Built on OpenBMB MiniCPM and DSpark, vLLM, AMD ROCm, AITER and Triton, with a huge shout out to to scorpsbr’s BF16 prefill work that increased the prefill on this model by up to 25%
Model Card:
huggingface.co/jcbtc/MiniCPM…
After seeing how good the new model is, I can't wait to try with the OpenBMB harness.
🚀 PilotDeck 2.0 is here!
Meet the open-source AI Agent Harness built to get complex work done. Developed by Tsinghua THUNLP, @modelbest_ai, @OpenBMB , and AI9Stars, PilotDeck is a model-agnostic Agent operating system that turns natural language instructions into completed tasks — with transparent data flows and the freedom to choose your models.
What’s new in 2.0?
📄 End-to-end office work: Create, edit, and verify Word, Excel, and PowerPoint files — not just generate drafts.
🔄 More reliable execution: complex, long-running tasks going through interruptions. Stress-tested with 2,000 interaction messages.
🛠️ Easy to get started: One-click desktop installation and a streamlined setup flow to connect your model and create your workspace.
🔓 Open by design: Fully open-source, model-agnostic, with data flows visible to users.
On Workbuddy Bench, PilotDeck scored 79.1 with Qwen3.6-35B-A3B, outperforming WorkBuddy (74.7) and Claude Code (77.3).
Build your own AI workspace for the Agent era.
🔗 GitHub: github.com/OpenBMB/PilotDeck
Web: pilotdeck.openbmb.cn
Ciru.ai - Crown 👑 retweeted
Get your @AMD Ryzen AI Max+ now!
AMD's Gorgon Halo is so good it's scary! Don't worry though, this AI monster wont turn you to stone 😂 But with 50% more RAM than last generation you can run som massive models on it! Learn more on Micro Center News: micro.center/u9kv
ALT A sleek, rectangular device with black, textured sides displaying the "AMD" logo. The device rests on a reflective surface, set against a blurred background with blue tones.
Ciru.ai - Crown 👑 retweeted
That's why I bought the Strix Halo in the first place. I love the local models I have (thanks to you). However, I knew that my frontier subs would be orchestrating to them. That said, I wanted a device that could handle practical workloads for me outside of just running models. The strix has been absolutely perfect for this.
I'm extremely impressed by this model so far.
@openbnb have done an incredible job.
I can tell you right now it's not benchmaxed.
I've been testing it the last 2 days and it's improving everything that I put it in.
I'm going to be releasing more soon.
🎬 35B MoE multimodal AI. On your phone. With just 3GB of memory. 🤯
This is MiniCPM-V 4.7 in action: reading real-world signs, planning around constraints, parsing complex documents, and reasoning over visual data.
Powered by LES (Lookahead Expert Scheduling), the 35B-A3B model brings efficient MoE inference to mobile devices, while the lightweight 1B variant delivers capable multimodal understanding on resource-constrained hardware.
🛠️ With Self-Replay Fine-Tuning (SRFT), developers can customize models for new tasks while maintaining general performance.
From iOS to Android, MiniCPM-V 4.7 makes powerful multimodal AI more practical to run locally.
Try it 👇
🔗 github.com/OpenBMB/MiniCPM-V
🤗 huggingface.co/collections/o…
Self-Replay Fine-Tuning: openbmb.cn/en/blog/self-repl…
Hear me out.
Sell your Spark clusters NOW.
Do not listen to the hype.
The things coming in the next 6 months will make that machine look like an overpriced toy.
Intel, AMD, Apple, and even smaller Chinese firms are about to dominate on price and performance. There will be no reason to even think of owning a spark at current prices. They are not good general machines either where the competition like AMD Ryzen AI Halo or Mac Studio are whole PCs that make incredible work machines in addition to providing comparable or superior inference.
This is very interesting. Great to see competition in local AI hardware.
Between this and apples new studio and all of the things AMD is doing I wouldn't want to be holding a spark cluster right now.
The same people using cloud models all day are the same ones who swear local AI is unusable at anything less than 100 tps.
The one who brought turboquant to local AI @no_stp_on_snek has now made @Italianclownz rocmfpx flash model competitive with some of the best.
Really impressive work.
More details and benchmarks here:
llm.ciru.ai/research/tom-tur…
A
Did I mention I love padogas?
My 3.8 flash orca build with default padoga prompt is making Qwen 3.8 max quality builds.
llm.ciru.ai/pixeltest
Had to copy Tom's cool scene.
Arche is orchestrating the fleet.
Dunamis is creating videos for everyone.
Ciru is working on a new q8 ornith build
Sozo is alternating between padogas and a new AMD demo I'm working on
Heptakis has been grinding on a project for @bensig for a week
The AI lab kitchen is cooking this morning.
M5 Max 128 - polishing qwen 3.8 27b speeds
5090 - working on gemma 31b integration
3090 - working on gemma 26b 14b integration
9070xt - working on gemma 12b integration
Dual Asus Sparks - speed improvements for qwen 3.8 flash next.
Mini m2 pro 32 - CI runner and kernel replay tester
@waffuruai
Ciru.ai - Crown 👑 retweeted
@boxwrenchx is single-handed bring AMD support to Maya!! What a God! That is what community is about 🔥
The remaining merge should be completed within the hour 🫡
Preliminary AMD (ROCm) work on Maya: GLM-5.3-Flash, Maya-S quant, Linux, one GPU
RX 7900 XT: ~413 tok/s prefill, ~16 tok/s decode
Radeon AI PRO R9700: ~490-560 prefill, ~20 decode
Strix Halo (8060S): ~217 prefill, ~18 decode
PRs:
github.com/mw00/project-maya… (merged)
github.com/mw00/project-maya…
